kernel/sched/wait.c

   1 /*
   2  * Generic waiting primitives.
   3  *
   4  * (C) 2004 Nadia Yvette Chambers, Oracle
   5  */
   6 #include "sched.h"
   7
   8 void __init_waitqueue_head(struct wait_queue_head *wq_head, const char *name, struct lock_class_key *key)
   9 {
  10         spin_lock_init(&wq_head->lock);
  11         lockdep_set_class_and_name(&wq_head->lock, key, name);
  12         INIT_LIST_HEAD(&wq_head->head);
  13 }
  14
  15 EXPORT_SYMBOL(__init_waitqueue_head);
  16
  17 void add_wait_queue(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
  18 {
  19         unsigned long flags;
  20
  21         wq_entry->flags &= ~WQ_FLAG_EXCLUSIVE;
  22         spin_lock_irqsave(&wq_head->lock, flags);
  23         __add_wait_queue(wq_head, wq_entry);
  24         spin_unlock_irqrestore(&wq_head->lock, flags);
  25 }
  26 EXPORT_SYMBOL(add_wait_queue);
  27
  28 void add_wait_queue_exclusive(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
  29 {
  30         unsigned long flags;
  31
  32         wq_entry->flags |= WQ_FLAG_EXCLUSIVE;
  33         spin_lock_irqsave(&wq_head->lock, flags);
  34         __add_wait_queue_entry_tail(wq_head, wq_entry);
  35         spin_unlock_irqrestore(&wq_head->lock, flags);
  36 }
  37 EXPORT_SYMBOL(add_wait_queue_exclusive);
  38
  39 void remove_wait_queue(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
  40 {
  41         unsigned long flags;
  42
  43         spin_lock_irqsave(&wq_head->lock, flags);
  44         __remove_wait_queue(wq_head, wq_entry);
  45         spin_unlock_irqrestore(&wq_head->lock, flags);
  46 }
  47 EXPORT_SYMBOL(remove_wait_queue);
  48
  49 /*
  50  * Scan threshold to break wait queue walk.
  51  * This allows a waker to take a break from holding the
  52  * wait queue lock during the wait queue walk.
  53  */
  54 #define WAITQUEUE_WALK_BREAK_CNT 64
  55
  56 /*
  57  * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
  58  * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
  59  * number) then we wake all the non-exclusive tasks and one exclusive task.
  60  *
  61  * There are circumstances in which we can try to wake a task which has already
  62  * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
  63  * zero in this (rare) case, and we handle it by continuing to scan the queue.
  64  */
  65 static int __wake_up_common(struct wait_queue_head *wq_head, unsigned int mode,
  66                         int nr_exclusive, int wake_flags, void *key,
  67                         wait_queue_entry_t *bookmark)
  68 {
  69         wait_queue_entry_t *curr, *next;
  70         int cnt = 0;
  71
  72         lockdep_assert_held(&wq_head->lock);
  73
  74         if (bookmark && (bookmark->flags & WQ_FLAG_BOOKMARK)) {
  75                 curr = list_next_entry(bookmark, entry);
  76
  77                 list_del(&bookmark->entry);
  78                 bookmark->flags = 0;
  79         } else
  80                 curr = list_first_entry(&wq_head->head, wait_queue_entry_t, entry);
  81
  82         if (&curr->entry == &wq_head->head)
  83                 return nr_exclusive;
  84
  85         list_for_each_entry_safe_from(curr, next, &wq_head->head, entry) {
  86                 unsigned flags = curr->flags;
  87                 int ret;
  88
  89                 if (flags & WQ_FLAG_BOOKMARK)
  90                         continue;
  91
  92                 ret = curr->func(curr, mode, wake_flags, key);
  93                 if (ret < 0)
  94                         break;
  95                 if (ret && (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
  96                         break;
  97
  98                 if (bookmark && (++cnt > WAITQUEUE_WALK_BREAK_CNT) &&
  99                                 (&next->entry != &wq_head->head)) {
 100                         bookmark->flags = WQ_FLAG_BOOKMARK;
 101                         list_add_tail(&bookmark->entry, &next->entry);
 102                         break;
 103                 }
 104         }
 105
 106         return nr_exclusive;
 107 }
 108
 109 static void __wake_up_common_lock(struct wait_queue_head *wq_head, unsigned int mode,
 110                         int nr_exclusive, int wake_flags, void *key)
 111 {
 112         unsigned long flags;
 113         wait_queue_entry_t bookmark;
 114
 115         bookmark.flags = 0;
 116         bookmark.private = NULL;
 117         bookmark.func = NULL;
 118         INIT_LIST_HEAD(&bookmark.entry);
 119
 120         spin_lock_irqsave(&wq_head->lock, flags);
 121         nr_exclusive = __wake_up_common(wq_head, mode, nr_exclusive, wake_flags, key, &bookmark);
 122         spin_unlock_irqrestore(&wq_head->lock, flags);
 123
 124         while (bookmark.flags & WQ_FLAG_BOOKMARK) {
 125                 spin_lock_irqsave(&wq_head->lock, flags);
 126                 nr_exclusive = __wake_up_common(wq_head, mode, nr_exclusive,
 127                                                 wake_flags, key, &bookmark);
 128                 spin_unlock_irqrestore(&wq_head->lock, flags);
 129         }
 130 }
 131
 132 /**
 133  * __wake_up - wake up threads blocked on a waitqueue.
 134  * @wq_head: the waitqueue
 135  * @mode: which threads
 136  * @nr_exclusive: how many wake-one or wake-many threads to wake up
 137  * @key: is directly passed to the wakeup function
 138  *
 139  * If this function wakes up a task, it executes a full memory barrier before
 140  * accessing the task state.
 141  */
 142 void __wake_up(struct wait_queue_head *wq_head, unsigned int mode,
 143                         int nr_exclusive, void *key)
 144 {
 145         __wake_up_common_lock(wq_head, mode, nr_exclusive, 0, key);
 146 }
 147 EXPORT_SYMBOL(__wake_up);
 148
 149 /*
 150  * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
 151  */
 152 void __wake_up_locked(struct wait_queue_head *wq_head, unsigned int mode, int nr)
 153 {
 154         __wake_up_common(wq_head, mode, nr, 0, NULL, NULL);
 155 }
 156 EXPORT_SYMBOL_GPL(__wake_up_locked);
 157
 158 void __wake_up_locked_key(struct wait_queue_head *wq_head, unsigned int mode, void *key)
 159 {
 160         __wake_up_common(wq_head, mode, 1, 0, key, NULL);
 161 }
 162 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
 163
 164 void __wake_up_locked_key_bookmark(struct wait_queue_head *wq_head,
 165                 unsigned int mode, void *key, wait_queue_entry_t *bookmark)
 166 {
 167         __wake_up_common(wq_head, mode, 1, 0, key, bookmark);
 168 }
 169 EXPORT_SYMBOL_GPL(__wake_up_locked_key_bookmark);
 170
 171 /**
 172  * __wake_up_sync_key - wake up threads blocked on a waitqueue.
 173  * @wq_head: the waitqueue
 174  * @mode: which threads
 175  * @nr_exclusive: how many wake-one or wake-many threads to wake up
 176  * @key: opaque value to be passed to wakeup targets
 177  *
 178  * The sync wakeup differs that the waker knows that it will schedule
 179  * away soon, so while the target thread will be woken up, it will not
 180  * be migrated to another CPU - ie. the two threads are 'synchronized'
 181  * with each other. This can prevent needless bouncing between CPUs.
 182  *
 183  * On UP it can prevent extra preemption.
 184  *
 185  * If this function wakes up a task, it executes a full memory barrier before
 186  * accessing the task state.
 187  */
 188 void __wake_up_sync_key(struct wait_queue_head *wq_head, unsigned int mode,
 189                         int nr_exclusive, void *key)
 190 {
 191         int wake_flags = 1; /* XXX WF_SYNC */
 192
 193         if (unlikely(!wq_head))
 194                 return;
 195
 196         if (unlikely(nr_exclusive != 1))
 197                 wake_flags = 0;
 198
 199         __wake_up_common_lock(wq_head, mode, nr_exclusive, wake_flags, key);
 200 }
 201 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
 202
 203 /*
 204  * __wake_up_sync - see __wake_up_sync_key()
 205  */
 206 void __wake_up_sync(struct wait_queue_head *wq_head, unsigned int mode, int nr_exclusive)
 207 {
 208         __wake_up_sync_key(wq_head, mode, nr_exclusive, NULL);
 209 }
 210 EXPORT_SYMBOL_GPL(__wake_up_sync);      /* For internal use only */
 211
 212 void __wake_up_pollfree(struct wait_queue_head *wq_head)
 213 {
 214         __wake_up(wq_head, TASK_NORMAL, 0, poll_to_key(EPOLLHUP | POLLFREE));
 215         /* POLLFREE must have cleared the queue. */
 216         WARN_ON_ONCE(waitqueue_active(wq_head));
 217 }
 218
 219 /*
 220  * Note: we use "set_current_state()" _after_ the wait-queue add,
 221  * because we need a memory barrier there on SMP, so that any
 222  * wake-function that tests for the wait-queue being active
 223  * will be guaranteed to see waitqueue addition _or_ subsequent
 224  * tests in this thread will see the wakeup having taken place.
 225  *
 226  * The spin_unlock() itself is semi-permeable and only protects
 227  * one way (it only protects stuff inside the critical region and
 228  * stops them from bleeding out - it would still allow subsequent
 229  * loads to move into the critical region).
 230  */
 231 void
 232 prepare_to_wait(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry, int state)
 233 {
 234         unsigned long flags;
 235
 236         wq_entry->flags &= ~WQ_FLAG_EXCLUSIVE;
 237         spin_lock_irqsave(&wq_head->lock, flags);
 238         if (list_empty(&wq_entry->entry))
 239                 __add_wait_queue(wq_head, wq_entry);
 240         set_current_state(state);
 241         spin_unlock_irqrestore(&wq_head->lock, flags);
 242 }
 243 EXPORT_SYMBOL(prepare_to_wait);
 244
 245 void
 246 prepare_to_wait_exclusive(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry, int state)
 247 {
 248         unsigned long flags;
 249
 250         wq_entry->flags |= WQ_FLAG_EXCLUSIVE;
 251         spin_lock_irqsave(&wq_head->lock, flags);
 252         if (list_empty(&wq_entry->entry))
 253                 __add_wait_queue_entry_tail(wq_head, wq_entry);
 254         set_current_state(state);
 255         spin_unlock_irqrestore(&wq_head->lock, flags);
 256 }
 257 EXPORT_SYMBOL(prepare_to_wait_exclusive);
 258
 259 void init_wait_entry(struct wait_queue_entry *wq_entry, int flags)
 260 {
 261         wq_entry->flags = flags;
 262         wq_entry->private = current;
 263         wq_entry->func = autoremove_wake_function;
 264         INIT_LIST_HEAD(&wq_entry->entry);
 265 }
 266 EXPORT_SYMBOL(init_wait_entry);
 267
 268 long prepare_to_wait_event(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry, int state)
 269 {
 270         unsigned long flags;
 271         long ret = 0;
 272
 273         spin_lock_irqsave(&wq_head->lock, flags);
 274         if (unlikely(signal_pending_state(state, current))) {
 275                 /*
 276                  * Exclusive waiter must not fail if it was selected by wakeup,
 277                  * it should "consume" the condition we were waiting for.
 278                  *
 279                  * The caller will recheck the condition and return success if
 280                  * we were already woken up, we can not miss the event because
 281                  * wakeup locks/unlocks the same wq_head->lock.
 282                  *
 283                  * But we need to ensure that set-condition + wakeup after that
 284                  * can't see us, it should wake up another exclusive waiter if
 285                  * we fail.
 286                  */
 287                 list_del_init(&wq_entry->entry);
 288                 ret = -ERESTARTSYS;
 289         } else {
 290                 if (list_empty(&wq_entry->entry)) {
 291                         if (wq_entry->flags & WQ_FLAG_EXCLUSIVE)
 292                                 __add_wait_queue_entry_tail(wq_head, wq_entry);
 293                         else
 294                                 __add_wait_queue(wq_head, wq_entry);
 295                 }
 296                 set_current_state(state);
 297         }
 298         spin_unlock_irqrestore(&wq_head->lock, flags);
 299
 300         return ret;
 301 }
 302 EXPORT_SYMBOL(prepare_to_wait_event);
 303
 304 /*
 305  * Note! These two wait functions are entered with the
 306  * wait-queue lock held (and interrupts off in the _irq
 307  * case), so there is no race with testing the wakeup
 308  * condition in the caller before they add the wait
 309  * entry to the wake queue.
 310  */
 311 int do_wait_intr(wait_queue_head_t *wq, wait_queue_entry_t *wait)
 312 {
 313         if (likely(list_empty(&wait->entry)))
 314                 __add_wait_queue_entry_tail(wq, wait);
 315
 316         set_current_state(TASK_INTERRUPTIBLE);
 317         if (signal_pending(current))
 318                 return -ERESTARTSYS;
 319
 320         spin_unlock(&wq->lock);
 321         schedule();
 322         spin_lock(&wq->lock);
 323
 324         return 0;
 325 }
 326 EXPORT_SYMBOL(do_wait_intr);
 327
 328 int do_wait_intr_irq(wait_queue_head_t *wq, wait_queue_entry_t *wait)
 329 {
 330         if (likely(list_empty(&wait->entry)))
 331                 __add_wait_queue_entry_tail(wq, wait);
 332
 333         set_current_state(TASK_INTERRUPTIBLE);
 334         if (signal_pending(current))
 335                 return -ERESTARTSYS;
 336
 337         spin_unlock_irq(&wq->lock);
 338         schedule();
 339         spin_lock_irq(&wq->lock);
 340
 341         return 0;
 342 }
 343 EXPORT_SYMBOL(do_wait_intr_irq);
 344
 345 /**
 346  * finish_wait - clean up after waiting in a queue
 347  * @wq_head: waitqueue waited on
 348  * @wq_entry: wait descriptor
 349  *
 350  * Sets current thread back to running state and removes
 351  * the wait descriptor from the given waitqueue if still
 352  * queued.
 353  */
 354 void finish_wait(struct wait_queue_head *wq_head, struct wait_queue_entry *wq_entry)
 355 {
 356         unsigned long flags;
 357
 358         __set_current_state(TASK_RUNNING);
 359         /*
 360          * We can check for list emptiness outside the lock
 361          * IFF:
 362          *  - we use the "careful" check that verifies both
 363          *    the next and prev pointers, so that there cannot
 364          *    be any half-pending updates in progress on other
 365          *    CPU's that we haven't seen yet (and that might
 366          *    still change the stack area.
 367          * and
 368          *  - all other users take the lock (ie we can only
 369          *    have _one_ other CPU that looks at or modifies
 370          *    the list).
 371          */
 372         if (!list_empty_careful(&wq_entry->entry)) {
 373                 spin_lock_irqsave(&wq_head->lock, flags);
 374                 list_del_init(&wq_entry->entry);
 375                 spin_unlock_irqrestore(&wq_head->lock, flags);
 376         }
 377 }
 378 EXPORT_SYMBOL(finish_wait);
 379
 380 int autoremove_wake_function(struct wait_queue_entry *wq_entry, unsigned mode, int sync, void *key)
 381 {
 382         int ret = default_wake_function(wq_entry, mode, sync, key);
 383
 384         if (ret)
 385                 list_del_init(&wq_entry->entry);
 386
 387         return ret;
 388 }
 389 EXPORT_SYMBOL(autoremove_wake_function);
 390
 391 static inline bool is_kthread_should_stop(void)
 392 {
 393         return (current->flags & PF_KTHREAD) && kthread_should_stop();
 394 }
 395
 396 /*
 397  * DEFINE_WAIT_FUNC(wait, woken_wake_func);
 398  *
 399  * add_wait_queue(&wq_head, &wait);
 400  * for (;;) {
 401  *     if (condition)
 402  *         break;
 403  *
 404  *     // in wait_woken()                       // in woken_wake_function()
 405  *
 406  *     p->state = mode;                         wq_entry->flags |= WQ_FLAG_WOKEN;
 407  *     smp_mb(); // A                           try_to_wake_up():
 408  *     if (!(wq_entry->flags & WQ_FLAG_WOKEN))     <full barrier>
 409  *         schedule()                              if (p->state & mode)
 410  *     p->state = TASK_RUNNING;                       p->state = TASK_RUNNING;
 411  *     wq_entry->flags &= ~WQ_FLAG_WOKEN;       ~~~~~~~~~~~~~~~~~~
 412  *     smp_mb(); // B                           condition = true;
 413  * }                                            smp_mb(); // C
 414  * remove_wait_queue(&wq_head, &wait);          wq_entry->flags |= WQ_FLAG_WOKEN;
 415  */
 416 long wait_woken(struct wait_queue_entry *wq_entry, unsigned mode, long timeout)
 417 {
 418         /*
 419          * The below executes an smp_mb(), which matches with the full barrier
 420          * executed by the try_to_wake_up() in woken_wake_function() such that
 421          * either we see the store to wq_entry->flags in woken_wake_function()
 422          * or woken_wake_function() sees our store to current->state.
 423          */
 424         set_current_state(mode); /* A */
 425         if (!(wq_entry->flags & WQ_FLAG_WOKEN) && !is_kthread_should_stop())
 426                 timeout = schedule_timeout(timeout);
 427         __set_current_state(TASK_RUNNING);
 428
 429         /*
 430          * The below executes an smp_mb(), which matches with the smp_mb() (C)
 431          * in woken_wake_function() such that either we see the wait condition
 432          * being true or the store to wq_entry->flags in woken_wake_function()
 433          * follows ours in the coherence order.
 434          */
 435         smp_store_mb(wq_entry->flags, wq_entry->flags & ~WQ_FLAG_WOKEN); /* B */
 436
 437         return timeout;
 438 }
 439 EXPORT_SYMBOL(wait_woken);
 440
 441 int woken_wake_function(struct wait_queue_entry *wq_entry, unsigned mode, int sync, void *key)
 442 {
 443         /* Pairs with the smp_store_mb() in wait_woken(). */
 444         smp_mb(); /* C */
 445         wq_entry->flags |= WQ_FLAG_WOKEN;
 446
 447         return default_wake_function(wq_entry, mode, sync, key);
 448 }
 449 EXPORT_SYMBOL(woken_wake_function);